Service 03 — Fine-Tuning & Private Models
Turn your enterprise data into specialised AI.
Generic models do not understand your business, your domain, your workflows or your terminology. We fine-tune open-source and foundation models for specific enterprise use cases — and deploy the result into your own cloud or on-premises infrastructure.
Deployment scope
On-premises / private VPC
Overview
We fine-tune open-source and foundation models on your domain-specific workflows and terminology using SFT with LoRA or QLoRA, correcting behaviour further with DPO where supervised training lands close but not quite right. The resulting model is deployed privately, keeping sensitive data inside controlled environments throughout.
We are equally likely to tell you fine-tuning is the wrong answer. It teaches behaviour, not facts — if you need the model to know current information, retrieval is the right tool. Most engagements combine both: a tuned model that follows your format and domain conventions, retrieving from an index for facts.
Capabilities
How we deliver
Baseline first, always
We evaluate the untuned model with a strong prompt before training anything. 'Fine-tuning beat a weak prompt' is not a result, and the baseline is what tells you whether any of this work is necessary.
Data curation over data volume
Results are dominated by data quality. Consistent, correctly formatted, representative examples — including edge cases and refusals, which are where teams most often omit coverage and where production most often breaks. We curate aggressively and hold out a split from real production traffic rather than a synthetic benchmark.
LoRA by default, full fine-tuning by exception
Adapters train 0.1 to 2% of parameters, which is what makes fine-tuning of large models practical on a single node. Full fine-tuning is justified for substantial capability shifts or continued pretraining, not as a default.
Serving adapters, not merged weights
Base weights stay immutable and adapters are served separately, so many fine-tuned models can share one set of base weights. That preserves hot-swapping, multi-tenancy and a fast rollback path — none of which survive merging.
What you get
- A model that follows your format and domain conventions consistently.
- Sensitive training data and resulting weights kept inside your perimeter.
- A held-out evaluation set and harness you keep and can re-run.
- Versioning across base model, adapter, data and evaluation results.
- Serving with adapter hot-swapping and a fast rollback path.
Frequently asked questions
Can we fine-tune models on our proprietary data?
Yes. We specialise in SFT and alignment (DPO) of open-source models on your domain-specific workflows and terminology, with the resulting model deployed privately so your data is never exposed.
How much data do we need?
For format and style adaptation, a few hundred well-chosen examples often suffice. For meaningful behavioural change, low thousands. Beyond that, returns diminish sharply without a corresponding rise in data quality.
Should we fine-tune or use RAG?
RAG for facts, anything current, or when you need source attribution. Fine-tuning for consistent format, consistent style, domain-specific behaviour, or a smaller model that matches a larger one on a narrow task. Most production systems use both.
How long does a fine-tuning engagement take?
A LoRA run over a few thousand examples is typically hours on a single high-memory GPU. The engagement timeline is dominated by data curation and evaluation rather than the training run itself.
Do we own the resulting weights?
Yes. The adapter and any fine-tuned weights are yours, hosted on infrastructure you operate. Merged or quantised variants can be produced on request.
Talk to an infrastructure architect
Bring your traffic profile, data constraints and cost targets. We will tell you what is feasible and what it would take.
Book a consultation